幸福账单

A |     根据麦肯锡（McKinsey）今年稍早发布的测算，全球数据中心容量将从2025年的82吉瓦增至2030年的219吉瓦，相当于现在的2.7倍左右。    

         
On the blue track inside the National Speed Skating Oval, robots chased each other in the 400-meter race. A sprinting Tien Kung robot was tripped by a rival robot lying on the track but did not fall. Instead, it adjusted its posture in seconds and continued toward the finish line with a result of 38.15 seconds, breaking human 400-meter race world record of 43.03 seconds, held by South African athlete Wayde van Niekerk. 
This scene unfolded on Sunday, the first official competition day of the 2nd World Humanoid Robot Games (WHRG). Later that evening, a humanoid robot named Lightning, developed by Honor Co., broke the long-standing human 1,500-meter world record of 3:26 set by Moroccan athlete Hicham El Guerrouj with a time of 2:30.22. Just one night earlier, during the opening ceremony's 100-meter sprint and high jump demonstrations, humanoid robots had already shattered records long held by human athletes Usain Bolt and Javier Sotomayor — merely an appetizer for the five-day sci-fi spectacle.
The Games have attracted 666 teams from 16 countries across six continents, up 138 percent from its inaugural edition. The number of participating robots quadrupled to 2,056, the Global Times learned from the organizing committee.
Yet the biggest change this year was not only the scale, but also the technological maturity. More teams have shifted from remote-controlled robots to fully autonomous systems. Chinese observers, after watching several events, said humanoid robots are no longer limited to simple demonstrations of walking or dancing. Increasingly, they are being tested in environments that demand speed, coordination, perception and decision-making.
From movement to intelligence

Behind the record-breaking performances are a series of technological upgrades. During the running events, Global Times reporters observed that robots were able to make subtle movements around their hip joints while walking, allowing them to maintain balance and produce movements closer to human biomechanics.
Guo Yijie, technical director of Tien Kung Ultra, said the improvement came from advances in hardware design.
"Robot joints have improved in torque and speed. As competitions demand higher power output, the reliability of the main control board, power board and high-power battery discharge capability have become increasingly important," Guo told the Global Times.
According to Guo, the same Tien Kung Ultra robot was used in both sprinting and jumping events. For running, the waist needs rotation to balance the inertia generated by leg movements. For jumping, the upper body requires pitching movement to help the robot achieve greater distance or height, Guo said, noting that these can all be achieved with only slight adjustments to the waist configuration were needed, and it highly demonstrated to the robot design's flexibility.
Besides the hardware upgrades, developments in robots' faster decision-making in "brains' is better showcased in football competitions.
The Booster T2 robot used by many domestic and overseas football teams has increased its onboard computing capability by 10 times compared with the previous generation. The time from recognizing the goal to deciding whether to pass or shoot has been reduced to around 100 milliseconds - faster than the time humans usually need to blink.
Cheng Hao, a technical representative from Boosters Robotics, said the improvement represents an "end-to-end" transformation in embodied intelligence.
"In the past, robots relied on separate steps: perception, planning, decision-making and execution. Each stage was disconnected. Now, with an end-to-end motion foundation model, visual input can directly generate joint control commands," Cheng said.
The newly introduced table tennis and tennis competitions are also testing another important capability: perception.
"Robots are moving from simply recognizing the three-dimensional position of a ball to understanding the movement information of opponents and more complex environments," said Li Yinghui, a postdoc researcher and the leader from the Hong Kong University team at the table tennis event. Such capabilities developed in sports scenarios could eventually be transferred to other scenarios in industries or daily life, Li said. 
Beyond entertainment
Though clips from the robot competitions have made a buzz on social media, with some critics from Western media questioning why robots need to dance, play football or compete in athletics while arguing that these activities appear far from everyday needs.
Zhou Changjiu, Vice President of the RoboCup Federation and the Robotics Competition Association of the Asia-Pacific (RCAP), stressed robot sports are not designed for entertainment alone. "Robot competitions are testing the capabilities needed for humanoid robots to enter the real world," Zhou told the Global Times.
Different events evaluate different abilities. Sprinting tests speed, marathons examine battery endurance and energy management, while football provides a comprehensive test because of its dynamic and unpredictable nature, testing intelligence, manipulation, interaction, safety, cost and trust at the same time, Zhou said.
Unlike a single-task demonstration, football requires robots to cooperate. Players must understand their own roles, recognize teammates and opponents, and make collective decisions - a form of swarm intelligence that could become essential when robots work together in homes, factories and emergency rescue environments.
Entertainment, however, can generate business opportunities. Gu Jinghai, brand director of EYOU Robot Co, told the Global Times that the company sold 95,000 integrated joints in China in 2025 and expects sales to exceed 1 million units in 2026. Growth mainly comes from planetary joints used in entertainment robots, while harmonic joints for industrial robots are also expanding steadily, he said. 
Jiao Jichao, an executive from UBTECH Robotics, said the criticism that "Chinese robots can only sing and dance" reflects only a temporary stage of development. "Performance and dancing are also real commercial demands, and their existence has value," Jiao said. "Beyond entertainment, Chinese robots are already demonstrating intelligent capabilities that support practical applications in manufacturing, and partnerships with companies including Huawei and Foxconn, where robots have already entered production lines in small-scale applications."
The Outline of the 15th Five-Year Plan for National Economic and Social Development of China stated that the country will construct a comprehensive cultivation system for future industries including the embody intelligence as new economic growth drivers.
Zhou said the development of humanoid robots is not only about individual machines, but about building an entire ecosystem.
China's advantage lies in its complete supply chain, abundant application scenarios and manufacturing capabilities, while the fast growing industry still faces challenges such as the lack of unified standards.
Gu said the company has launched mass production of robotics joint under domestic standard of auto parts. 
Robot competitions can help address these issues by creating common standards in areas such as competition rules, robot dimensions, communication systems, batteries and joint modules. "As applications expand, the cost of key robot components will continue to decline," Zhou said. 
"Numbers are magic," Zhou said. "If producing one robot costs a huge amount because it is handmade, producing one million robots will completely change the cost structure."
For humanoid robots, the race taking place inside the Ice Ribbon may not ultimately be about who crosses the finish line first. It may be about whether these machines can learn to navigate the far more complicated race outside the stadium - the race into the real world.
。其中美国占比约45%，约90至95吉瓦。英伟达（Nvidia）首席执行官黄仁勋近期估算，基于英伟达架构新建一吉瓦算力的成本可能很快达到800亿至1000亿美元。按此换算，未来五年全球新增产能对应的资本开支是万亿美元量级。

B |          与投入规模同步升温的，是对超大云厂商资本开支的质疑。

C | 作为全球芯片板块的先行指标，费城半导体指数（SOX）在过去一个月中下跌了13.38%。         不过，美国投资分析机构NDR首席主题策略师乔西克（Pat Tschosik）在最近一场研讨会上表示，数据中心建设的物理扩张不会因为质疑声而立即放缓，电力设备、变压器、散热系统等物理层建设因两到五年的交付周期已被锁定。他表示，“超级周期尚未走完”，但对训练层资本开支的容忍度，正在进入一个更审慎的阶段。

D |          乔西克在回应第一财经记者关于折旧的提问时说，即便把折旧成本算进去，结论也不会变。“真正重要的还是云销售增长的数字。

E | ”他认为，一旦资本开支增速持续跑赢云计算收入增速，“资本开支义警”（capex vigilantes）就会出现，用抛售给出市场信号。         从“训练主导”转向“推理主导”          当前，费城半导体指数（SOX）两年期相对标普500的超额收益，在今年6月已升至125%至150%区间，这是NDR认为的“泡沫观察区”。

F | 上一次触及是在2021年2月，同年12月SOX相对强度即见顶。         但从预租率来看，对数据中心的需求目前仍真实存在。

G | 据仲量联行（JLL）数据，北美数据中心空置率约为1%，且自2024年底以来维持在这一水平；美国在建产能中92%已经预租，欧洲这一比例为82%。乔西克认为，这组数字说明需求本身仍然存在，建设方也愿意继续投入，但审批流程和电网接入速度构成硬约束，这一瓶颈在欧洲更为明显。         美国银行分析师也认为，未来五年美国将需要超过230吉瓦（GW）的新发电容量，但受监管的公用事业公司预计仅能增加约93吉瓦的合格供应量，从而留下了超过100吉瓦的电力缺口。报告显示，在此期间，单是数据中心就可能增加大约125吉瓦的美国电力负荷，从而将2026年至2030年的整体电力需求推高至4.1%的复合年增长率。

H |          不过，从更细分的角度来看，乔西克认为，就整体产能建设而言，目前处于“中局”（middle innings）；但专门用于模型训练的产能建设，已经进入“后段”（later innings），从2027年起，新增容量将更多流向推理需求，而非训练需求。         “训练与推理对基础设施的需求曲线并不相同，当训练需求见顶时，围绕训练能力构建的资本开支逻辑将首先面临检验。”他解释说，值得反复追问的是，“这些产能真的用得上吗？推理需求真的会起来吗？”          乔西克提出的另一个变量是生产力兑现的时滞。他的团队测算显示，科技支出占美国国内生产总值（GDP）比重的变化，与生产力提升之间存在约20个季度、即五年左右的滞后关系。据此推算，当前围绕模型训练进行的巨额投入，其对应的生产力回报要到数年后才会显现。         这意味着，即便整个超级周期尚未走完，推理需求带来的下一轮增长仍会到来，数据中心资本开支仍可能在生产力兑现之前，先经历一轮调整。

I | 乔西克将这一调整的时间窗口大致设定在两年左右，快于20个季度的历史滞后期，逻辑是一旦云厂商销售增速率先放缓，市场不会等到生产力数据兑现才做出反应。“我的基准情形是，数据中心会经历一轮调整，直到推理端真正开始发力为止。         如何判断数据中心繁荣期的尾声          根据标普全球市场智库（S&P Global Market Intelligence）的数据，2025年数据中心的并购（M&A）交易额达到了约500亿美元，是前一年的两倍多。

J |          乔西克提出，本轮数据中心繁荣可能以三种历史脚本之一收场：供给过剩、需求下滑，或融资枯竭，分别对应1999年、2022年和2008年的教训。         他为每一种脚本给出了一组可以实时追踪的指标。供给过剩是1999年互联网泡沫留下的教训。乔西克列出的观测指标包括数据中心建设周期、空置率和GPU租赁价格。

K | 目前建设周期仍需两到五年，尚未出现缩短迹象；空置率如前所述维持在约1%的低位；仲量联行和世邦魏理仕（CBRE）发布的数据均未显示供给过剩的信号。         需求下滑是他称为“最值得对照”的2022年教训。彼时，随着企业削减云计算开支，云厂商销售增速放缓，而资本开支仍在加速，亚马逊和英伟达股价当年跌幅均超过50%。

L | 乔西克提出的判断阈值是：当资本开支增速超过销售增速的1.5倍，市场通常会做出负面反应，这一关系在微软身上同样成立。         “如果只能挑一两个指标来看，第一位是超大规模云厂商的云业务增速，第二位就是它们的资本开支指引，”他表示，“一旦这些公司开始下调指引，往往说明它们已经察觉到需求端出了问题。”          融资枯竭对应2008年金融危机的教训。乔西克关注的核心变量是资本开支占经营现金流的比例，亚马逊这一比例已接近100%，亚马逊、微软、谷歌等超大规模云厂商合计约为70%。         “我相信，一旦这个比例逼近甚至超过100%，市场对这些公司融资的态度就不会像现在这样友善。”乔西克提到，谷歌近期已经开始发债支持扩张。

M | 此外，科技行业高收益债利差、“新云”公司（如CoreWeave、甲骨文）的股价表现，以及OpenAI和Anthropic两家公司未来的上市时点，都是判断融资环境是否收紧的辅助信号，“这说明现在的市场环境可能已经是最好的时候了，所以它们才要上市。”          折旧是否被低估          值得关注的是，资本开支占经营现金流的比例，有可能掩盖了折旧对企业盈利的实际侵蚀。该指标所依据的经营现金流会将折旧加回，因此无法反映折旧在利润表层面造成的拖累，而如果单独测算折旧压力，数据中心繁荣的拐点是否会提前出现？          对于第一财经记者的提问，乔西克在回应中承认，这一质疑成立。“确实，我能理解这会被视为一种激进的会计处理方式。”但他表示，更重要的变量是企业是否会因为不愿耗尽经营现金流，而转向发债融资。         “让我们把注意力放在真正重要的因素上，”他说，市场会看到企业债务开始增加，一旦市场对新增债务作出负面反应，就意味着“资本纪律”正式回归。他将这一现象称为“资本开支义警”。在他看来，与其纠结折旧的会计处理方式，不如紧盯云销售增速与资本开支增速的相对关系，这才是真正决定市场情绪的变量。         这一问题延伸出另一层担忧，即GPU资产本身的折旧年限是否被低估。乔西克的回应是，训练阶段需要性能最强的芯片，但推理阶段对算力速度的要求更低，这意味着旧款GPU在被训练环节淘汰后，仍可以流入推理场景继续使用，加之二手GPU市场已经存在，实际使用寿命可能长于市场的悲观假设。

N | “我不认为这像有些人渲染的那样严重。         (本文来自第一财经)。